Life of a request

It can sometimes be confusing to figure out how to write a new feature, or debug an existing one. Let us try to follow a request through the Zulip codebase, and dive deep into how each part works.

We will use as our example the creation of users through the API, but we will also highlight how alternative requests are handled.

A request is sent to the server, and handled by nginx 

When Zulip is deployed in production, all requests go through nginx. For the most part we don’t need to know how this works, except for when it isn’t working. nginx does the first level of routing–deciding which application will serve the request (or deciding to serve the request itself for static content).

In development, tools/run-dev.py fills the role of nginx. Static files are in your Git checkout under static, and are served unminified.

Static files are served directly by nginx

Static files include JavaScript, css, static assets (like emoji, avatars), and user uploads (if stored locally and not on S3).

File not found errors (404) are served using a Django URL, so that we can use configuration variables (like whether the user is logged in) in the 404 error page.

location /static/ {
    alias /home/zulip/prod-static/;
    # Set a nonexistent path, so we just serve the nice Django 404 page.
    error_page 404 /django_static_404.html;
}

nginx routes other requests between Django and Tornado 

All our connected clients hold open long-polling connections so that they can receive events (messages, presence notifications, and so on) in real-time. Events are served by Zulip’s tornado application.

Nearly every other kind of request is served by the zerver Django application.

Here is the relevant nginx routing configuration.

Django routes the request to a view in urls.py files

There are various urls.py files throughout the server codebase, which are covered in more detail in the directory structure doc.

The main Zulip Django app is zerver. The routes are found in zproject/urls.py and zproject/legacy_urls.py.

There are HTML-serving, REST API, legacy, and webhook url patterns. We will look at how each of these types of requests are handled, and focus on how the REST API handles our user creation example.

Views serving HTML are internationalized by server path

If we look in zproject/urls.py, we can see something called i18n_urls. These urls show up in the address bar of the browser, and serve HTML.

For example, the /features page (preview here) gets translated in Chinese at zh-hans/features/ (preview here).

Note the zh-hans prefix–that url pattern gets added by i18n_patterns.

API endpoints use REST 

Our example is a REST API endpoint. It’s a PUT to /users.

With the exception of incoming webhooks (which we do not usually control the format of), legacy endpoints, and logged-out endpoints, Zulip uses REST for its API. This means that we use:

POST for creating something new where we don’t have a unique ID. Also used as a catch-all if no other verb is appropriate.
PUT for creating something for which we have a unique ID.
DELETE for deleting something
PATCH for updating or editing attributes of something.
GET to get something (read-only)
HEAD to check the existence of something to GET, without getting it; useful to check a link without downloading a potentially large link
OPTIONS (handled automatically, see more below)

Of these, PUT, DELETE, HEAD, OPTIONS, and GET are idempotent, which means that we can send the request multiple times and get the same state on the server. You might get a different response after the first request, as we like to give our clients an error so they know that no new change was made by the extra requests.

POST is not idempotent–if I send a message multiple times, Zulip will show my message multiple times. PATCH is special–it can be idempotent, and we like to write API endpoints in an idempotent fashion, as much as possible.

This cookbook and tutorial can be helpful if you are new to REST web applications.

PUT is only for creating new things

If you’re used to using PUT to update or modify resources, you might find our convention a little strange.

We use PUT to create resources with unique identifiers, POST to create resources without unique identifiers (like sending a message with the same content multiple times), and PATCH to modify resources.

In our example, create_user_backend uses PUT, because there’s a unique identifier, the user’s email.

OPTIONS

The OPTIONS method will yield the allowed methods.

This request: OPTIONS https://chat.zulip.org/api/v1/users yields a response with this HTTP header: Allow: PUT, GET

We can see this reflected in zproject/urls.py:

rest_path('users',
          GET=get_members_backend,
          PUT=create_user_backend),

In this way, the API is partially self-documenting.

Legacy endpoints are used by the web client

The endpoints from the legacy JSON API are written without REST in mind. They are used extensively by the web client, and use POST.

You can see them in zproject/legacy_urls.py.

Incoming webhook integrations may not be RESTful

Zulip endpoints that are called by other services for integrations have to conform to the service’s request format. They are likely to use only POST.

Django calls rest_dispatch for REST endpoints, and authenticates

For requests that correspond to a REST url pattern, Zulip configures its url patterns (see zerver/lib/rest.py) so that the action called is rest_dispatch. This method will authenticate the user, either through a session token from a cookie, or from an email:api-key string given via HTTP basic auth for API clients.

It will then look up what HTTP verb was used (GET, POST, etc) to make the request, and then figure out which view to show from that.

In our example,

GET=get_members_backend,
PUT=create_user_backend

are supplied as arguments to rest_path, along with the HTTPRequest. The request has the HTTP verb PUT, which rest_dispatch can use to find the correct view to show: zerver.views.users.create_user_backend.

The view will authorize the user, extract request variables, and validate them

This is covered in good detail in the writing views doc.

Results are given as JSON

Our API works on JSON requests and responses. Every API endpoint should raise JsonableError in the case of an error, which gives a JSON string:

{"result": "error", "msg": "<some error message>", "code": "BAD_REQUEST"}

in a Django HttpResponse object with a Content-Type of ‘application/json’.

To pass back data from the server to the calling client, in the event of a successfully handled request, we use json_success(request, data).

The request argument is a Django HttpRequest object. The data argument is a Python object which can be converted to a JSON string and has a default value of an empty Python dictionary.

Zulip stores additional metadata it has associated with that HTTP request in a RequestNotes object, which is primarily accessed in common code used in all requests (middleware, logging, rate limiting, etc.).

This will result in a JSON string:

{"result": "success", "msg": "", "data": {"var_name1": "var_value1", "var_name2": "var_value2"}}

with a HTTP 200 status and a Content-Type of ‘application/json’.

That’s it!